What is expensive about a robot
What makes a robot expensive is rarely the idea and almost always the repetition. Every change to the body demands new parts, new assembly and new wiring, and costs weeks in which nothing is learned. Anyone who develops a humanoid in metal and plastic rebuilds for every insight.
We reversed the order. Our robot stands about 78 centimetres tall, weighs 4.4 kilograms, has 38 actuators and two hands with five fingers each, and to this day no part of it has been manufactured. It exists as a design and as a simulation. In this form its body went through several versions; between the fifth and the seventh alone, limb lengths, masses and joint limits changed. None of these changes needed a workshop.
The condition: the simulation must not flatter
What a robot can do in simulation it can often do only in part once built, and the gap is usually wider for seeing and grasping than for walking. The gap arises wherever the simulation makes things easier for the robot than reality does, with motors without delay, sensors without noise and a robot that knows things it could never measure.
We therefore wrote down three rules. Under the first we model every joint as the actuator that will later be installed, with torque, backlash and delay taken from the data sheet until measurements on the real part exist. Under the second the robot gets only what its sensors deliver: the camera in its head where the head points, the inertial sensor with its drift, the joint encoders with their errors, the force sensors in the feet and the touch pads in the fingers. The third forbids fitting the world to an individual task. So we move no table closer and relax no test so that the robot passes it.
Why we withdrew our best numbers
On 2 October 2026 we audited the second rule and found it broken in several places. The rule was in force, but nothing had enforced it. The robot oriented itself with a range finder it does not have, and its gait knew the true heading of the body instead of the measured one.
The results that rested on this looked good. The robot opened a door in 95% of attempts and put something down in 88%. We archived these results and no longer count them as the state of the robot, nor the 289 training runs and 737 intermediate versions of the gait. Since then only what the robot learns with the senses it will have once built counts. A test keeps a list of every place where the controller reads data from the simulation. If a new one appears, the test fails until someone has checked that one of the robot’s sensors can deliver that data.
The decision wrote off the compute of 289 runs. We took it because a number that does not hold on the built robot says nothing about it. If that only shows after the build, the finished robot has to learn its controller from scratch.
What exists today
The design is complete. 107 parts for printing have been checked in CAD for printability and fit, the purchased parts are listed, each of the 38 actuators has its place and its axis, and a manual leads through assembly in 23 steps. The build plan is meant to ensure that the robot on the bench is the same one the simulation trains. Where a value cannot be matched, we measure the real one and carry it back into the simulation.
So far the robot as a whole has learned to walk, with the senses it will have once built, and it did not fall in 108 tested walks. On a separate test bench its hand holds an object in just under three quarters of the trials. In tasks that combine seeing, walking and grasping it does not yet reach the pass marks. The project page gives the measurements, and Our coding agent designs a humanoid robot in 24 days tells how the project came about.
What is open
Only the built robot will show whether the three rules suffice. Every simulation leaves something out, and what is missing is only partly known in advance. The plan therefore starts with a single printed finger on the test bench and carries every deviation between measurement and model into the simulation before the next part is made. Until then everything we say about this robot describes a simulation.